!3513 fix bert bug

Merge pull request !3513 from 周莉莉/bert

official/nlp/bert/README.md

@@ -269,7 +269,8 @@ For ner or classification task, schema file contains ["input_ids", "input_mask",
 
 For squad task, training: schema file contains ["start_positions", "end_positions", "input_ids", "input_mask", "segment_ids"], evaluation: schema file contains ["input_ids", "input_mask", "segment_ids"].
 
-`numRows` is the only option which could be set by user, other values must be set according to the dataset.
+`numRows` is the only option in schema file which could be set by user when dataset_format is tfrecord, other values must be set according to the dataset.
+`num_samlpes` is the only option in yaml file which could be set by user when dataset_format is mindrecord, other values must be set according to the dataset.
 
 For example, the schema file of cn-wiki-128 dataset for pretraining shows as follows:
 {

official/nlp/bert/README_CN.md

@@ -272,7 +272,8 @@ For ner or classification task, schema file contains ["input_ids", "input_mask",
 
 For squad task, training: schema file contains ["start_positions", "end_positions", "input_ids", "input_mask", "segment_ids"], evaluation: schema file contains ["input_ids", "input_mask", "segment_ids"].
 
-`numRows` is the only option which could be set by user, other values must be set according to the dataset.
+`numRows` is the only option in schema file which could be set by user when dataset_format is tfrecord, other values must be set according to the dataset.
+`num_samlpes` is the only option in yaml file which could be set by user when dataset_format is mindrecord, other values must be set according to the dataset.
 
 For example, the schema file of cn-wiki-128 dataset for pretraining shows as follows:
 {

official/nlp/bert/pretrain_config.yaml

@@ -32,6 +32,7 @@ save_checkpoint_num: 1
 data_dir: ''
 schema_dir: ''
 dataset_format: "mindrecord"
+num_samples: None   # is the option which could be set by user to specify steps
 
 # ==============================================================================
 # pretrain related

official/nlp/bert/pretrain_config_Ascend_Boost.yaml

@@ -32,6 +32,7 @@ save_checkpoint_num: 1
 data_dir: ''
 schema_dir: ''
 dataset_format: "mindrecord"
+num_samples: None   # is the option which could be set by user to specify steps
 
 # ==============================================================================
 # pretrain related

official/nlp/bert/pretrain_config_Ascend_Thor.yaml

@@ -32,6 +32,7 @@ save_checkpoint_num: 5
 data_dir: ''
 schema_dir: ''
 dataset_format: "mindrecord"
+num_samples: None   # is the option which could be set by user to specify steps
 
 # ==============================================================================
 # pretrain related

official/nlp/bert/run_pretrain.py

@@ -231,7 +231,7 @@ def run_pretrain():
             logger.info("save checkpoint steps: {}".format(cfg.save_checkpoint_steps))
 
     ds = create_bert_dataset(device_num, rank, cfg.do_shuffle, cfg.data_dir, cfg.schema_dir, cfg.batch_size,
-                             cfg.bucket_list, cfg.dataset_format)
+                             cfg.bucket_list, cfg.dataset_format, cfg.num_samples)
     net_with_loss = BertNetworkWithLoss(bert_net_cfg, True)
 
     new_repeat_count = cfg.epoch_size * ds.get_dataset_size() // cfg.data_sink_steps
@@ -262,7 +262,7 @@ def run_pretrain():
     if cfg.train_with_eval == 'true':
         net_eval = BertPretrainEval(bert_net_cfg, network=net_with_loss.bert)
         eval_ds = create_eval_dataset(cfg.batch_size, device_num, rank, cfg.eval_data_dir, cfg.schema_dir,
-                                      cfg.dataset_format)
+                                      cfg.dataset_format, cfg.num_samples)
         model = Model(net_with_grads, eval_network=net_eval, metrics={'bert_acc': BertMetric(cfg.batch_size)})
         eval_callback = EvalCallBack(model, eval_ds, device_num * cfg.batch_size, cfg.eval_samples)
         callback.append(eval_callback)

official/nlp/bert/src/dataset.py

@@ -116,7 +116,7 @@ class BucketDatasetGenerator:
 
 
 def create_bert_dataset(device_num=1, rank=0, do_shuffle="true", data_dir=None, schema_dir=None, batch_size=32,
-                        bucket_list=None, dataset_format="mindrecord"):
+                        bucket_list=None, dataset_format="mindrecord", num_samples=None):
     """create train dataset"""
     # apply repeat operations
     files = os.listdir(data_dir)
@@ -126,11 +126,17 @@ def create_bert_dataset(device_num=1, rank=0, do_shuffle="true", data_dir=None,
                 (dataset_format == "mindrecord" and "mindrecord" in file_name and "mindrecord.db" not in file_name):
             data_files.append(os.path.join(data_dir, file_name))
     if dataset_format == "mindrecord":
-        data_set = ds.MindDataset(data_files,
-                                  columns_list=["input_ids", "input_mask", "segment_ids", "next_sentence_labels",
-                                                "masked_lm_positions", "masked_lm_ids", "masked_lm_weights"],
-                                  shuffle=ds.Shuffle.FILES if do_shuffle == "true" else False,
-                                  num_shards=device_num, shard_id=rank)
+        if num_samples is not None:
+            data_set = ds.MindDataset(data_files,
+                                      columns_list=["input_ids", "input_mask", "segment_ids", "next_sentence_labels",
+                                                    "masked_lm_positions", "masked_lm_ids", "masked_lm_weights"],
+                                      shuffle=False, num_shards=device_num, shard_id=rank, num_samples=num_samples)
+        else:
+            data_set = ds.MindDataset(data_files,
+                                      columns_list=["input_ids", "input_mask", "segment_ids", "next_sentence_labels",
+                                                    "masked_lm_positions", "masked_lm_ids", "masked_lm_weights"],
+                                      shuffle=ds.Shuffle.FILES if do_shuffle == "true" else False,
+                                      num_shards=device_num, shard_id=rank)
     elif dataset_format == "tfrecord":
         data_set = ds.TFRecordDataset(data_files, schema_dir if schema_dir != "" else None,
                                       columns_list=["input_ids", "input_mask", "segment_ids", "next_sentence_labels",
@@ -261,7 +267,7 @@ def create_squad_dataset(batch_size=1, data_file_path=None, schema_file_path=Non
 
 
 def create_eval_dataset(batchsize=32, device_num=1, rank=0, data_dir=None, schema_dir=None,
-                        dataset_format="mindrecord"):
+                        dataset_format="mindrecord", num_samples=None):
     """create evaluation dataset"""
     data_files = []
     if os.path.isdir(data_dir):
@@ -275,7 +281,8 @@ def create_eval_dataset(batchsize=32, device_num=1, rank=0, data_dir=None, schem
     if dataset_format == "mindrecord":
         data_set = ds.MindDataset(data_files,
                                   columns_list=["input_ids", "input_mask", "segment_ids", "next_sentence_labels",
-                                                "masked_lm_positions", "masked_lm_ids", "masked_lm_weights"])
+                                                "masked_lm_positions", "masked_lm_ids", "masked_lm_weights"],
+                                  num_samples=num_samples)
     elif dataset_format == "tfrecord":
         data_set = ds.TFRecordDataset(data_files, schema_dir if schema_dir != "" else None,
                                       columns_list=["input_ids", "input_mask", "segment_ids", "next_sentence_labels",

research/cv/DBPN/README.md

@@ -210,7 +210,7 @@ Inference result is saved in current path, you can find result like this in acc.
 | Resource                   | CentOs 8.2; Ascend 910; CPU 2.60GHz, 192cores; Memory 755G  |
 | MindSpore Version          | 1.3.0                                                       |
 | Dataset                    | DIV2K                                                       |
-| Training Parameters        | ddbpn:epoch=2000, batch_size = 16; dbpngan:epoch=1000,batch_size=4|
+| Training Parameters        | ddbpn:epoch=2000, batch_size = 16; dbpngan:epoch=1100,batch_size=4|
 | Optimizer                  | Adam                                                        |
 | Loss Function              | BCELoss  MSELoss VGGLoss                                    |
 | outputs                    | super-resolution pictures                                   |

research/cv/DBPN/src/util/config.py

@@ -52,7 +52,7 @@ def get_args(is_gan=False):
     # additional parameters
     parser.add_argument('--sens', type=float, default=1024.0)
     if is_gan:
-        parser.add_argument('--nEpochs', type=int, default=1000, help='number of epochs to train for')
+        parser.add_argument('--nEpochs', type=int, default=1100, help='number of epochs to train for')
         parser.add_argument('--batchSize', type=int, default=4, choices=[4, 8, 16], help='training batch size')
         parser.add_argument('--patch_size', type=int, default=60, choices=[40, 60], help='Size of cropped HR image')
         parser.add_argument('--pretrained_iter', type=int, default=100, help='number of epochs to train for')